# There are so many variables. You have the base model, but there’s also the scheduler and the sampler—plus the number of steps. You might even use a LoRA trained with a specific number of steps. When training a LoRA, you can choose settings like high, medium, or low noise bias ... — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/b28fb8306e27c466

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-11 20:47 ingested 2026-11-12 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xtt6zf/there_are_so_many_variables_you_have_the_base/)

## retained text



```
as well as shift, weighted, sigmoid, or linear approaches. Then there’s the training resolution—and on top of that, you also have the inference shift

LoRA number of steps, training rate, batch size, and simulated batch size

Could the problem be the model?

What about the LoRA?

The resolution?

A shift between training and inference?

Is low noise useless?

Should I run a few more steps, or fewer?

Samplers—there are at least 20 of them. And schedulers—more than 10.

Then there are other things, like ETA and noise injection.
```



document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtt6zf/there_are_so_many_variables_you_have_the_base
